Garth Sundem   00:00

Welcome to this episode of the Medical Affairs Professional Society podcast series: “Elevate”. I’m your host, Garth Sundem, Communications Director at MAPS. And today we’re talking about generative AI in Medical Affairs insights with experts from Virtual Science AI, including CEO Tom Hughes, Chief Customer Officer Geraldine Reilly, and Chief Technology Officer Han Chen. This episode is sponsored by Virtual Science AI. So, Tom, AI, gen AI, most of our listeners are interested in this space, and have heard of generative AI. But I’m not sure I know exactly what generative AI means. And I wonder if our listeners are in the same place. So could we pretty please start with What does generative AI mean? And what is it doing in Medical Affairs now?

Tom Hughes   01:03

Absolutely, it’s an excellent question golf. And it’s great to be here. And thanks for having us on the MAPS podcast. Let me start by saying Virtual Science AI is a Life Science Advisory Panel and former AI insight management solution company. You know, we’re the first in the life science industry to capture an aggregate all interactive data from advisory panels and produce rapid insight reports that gives information on trends, themes and topics coming out of those activities. We’ve developed medically trained AI that supports in this process. And we’re working with many of the top leading pharmaceutical companies and biotechs. So what is generative AI? And how is it being applied in Medical Affairs, I will hit it at a high level. And I’ll ask Hans to make a comment our chief technology officer on genitive AI more technically. And then we’ll back into Medical Affairs as a group and talk about what we’re seeing and use cases and how it’s improving outcomes. But from my view, generative AI is a type of AI, it’s a transformer model. And ultimately, how generative AI works is somewhat like a game. So the model is trained on certain datasets, and then it’s trying to best guess the next word. And then that model is being trained for different training techniques, such as reinforced learning, supervised learning, unsupervised learning, to help improve the accuracy of the model to serve the use case that that generative model is being designed for. I’m gonna ask Hans to talk more specifically to generative AI, and what it is, and then we’ll get to talk a bit more specifically about advancements in Medical Affairs, but the use of the Gen AI

Garth Sundem   03:24

Alright, Hans, before we get to that, let me just ask is the generative part that best guess so the thing is trained, however it’s trained? And then it guesses you said the next best word? Is that the difference? Or the definition of generative AI, compared to just the AI we’ve seen since? Oh, I don’t know. I mean, even the what 1980s or 90s? Is that creative?

Chih Han Chen   03:52

Yeah. So So that’s, that’s a very good way to look at it. So, generative AI, if you really look at the word generic, it is generating the next word generating what predicts the next word and and that actually allow a lot of potentials is called self to self learning that allows the goes way and we the articles are on from the public Internet, and then learn from these contents. And so, basically, you follow like the, you can imagine as a reading procedure trying to predict the next word and then learn from it, whether this next word is daiquiri or not. So that that is is generally AI usually means and then from another perspective, look at this is that J AI is predicting the next floods based on the content that previously feeling so in learning from the past experience, what has been trending.

Garth Sundem   04:56

You know that that’s such an interesting thing. You know, the word generative seems like, has really driven the explosion of AI into the public consciousness just now, you know, where it’s not just asking AI to find interesting things in a data set or things that we’ve done before, but it’s asking AI to generate new knowledge. And I know that’s I know, that’s a sticky way to frame it. Because it’s taking old things and synthesizing it, but it’s the generative aspect that really seems like Medical Affairs is now on board with leveraging. So where are we at them with the Medical Affairs uses of this generative AI?

Tom Hughes   05:43

Let me start by saying that, you know, from the many interactions that we have across the industry, and from the interactions that, you know, we have it, you know, MAPS with the community, you know, this is a hot topic, and there’s a lot of interest and a lot of appetite to experiment. And, you know, even at the most senior levels, you know, you know, AI is coming up in earnings calls and pharma companies are investing, you know, significant amounts of money into AI. And it’s somewhat Do or die when it comes to the use of AI. You know, there’s a perception, I believe that we’re starting to see companies raced to start to apply AI to outpace their competitors by being able to do things faster, or create innovation in ways that hasn’t been done before. Now, specific to Medical Affairs, where Medical Affairs has been challenged, is in how to apply AI to the function of Medical Affairs. And one of the biggest concerns and restrictions has been around the use of chat GPT in the complaint were, and many of the largest companies and biotechs put guidance out and say, do not use chat GPT. And the reason why they’ve done that is because those large language models are public lives language models. So if you put confidential and sensitive information into an LLM, like chat GPT or another LLM, you know this risk, that information could be shared with another company, and then they might get the learning benefit, and access to that sensitive and confidential information. And because of that, you know, it’s raised alarms, and it’s, it’s prevented companies to jump in and start to use these public facing LLM to apply it to Medical Affairs. But what’s, you know, we’re proud to say about Virtual Science AI is that we’ve designed pharma compliant AI, which allows teams to safely and compliantly work with AI in the function of Medical Affairs, because all models are unique to the client. And then those benefits are not shared between the other companies. And because of that were approved by many of the largest pharmaceutical companies in the world, and increasingly across borders. And, you know, companies are applying our AI, specifically to advisory panel data to get reports within 24 hours following activity that would have taken three to six weeks. You know, they’re able to get information and quantitative outputs, such as sentiment around key strategic topics. You know, companies are also looking to go beyond the advisory board data and start to ingest over data, such as pre and post news around medic congresses, and try to transcribe stakeholder responses. You know, there’s different use cases that are coming up that we’re helping with to start to make sense of complex medical information to deliver more fast, more meaningful and actionable insights.

Garth Sundem   09:31

Well, okay, so, these large language models, they try GPT you know, they’re Chang trained on the Internet, what pre 2021 And because they have such a large data set, they can be very accurate, but they can also be sort of pulled one way or the other by everything that is out there. Do you see an advantage Have more medically trained Gen AIS, or or AI is large language models, Gen AI models that aren’t trained on the full internet. But but are trained more on medical data. Maybe that’s one company maybe that’s many what what does that mean? Yeah,

Tom Hughes   10:21

it’s a really interesting debate in the field of AI LLM ‘s versus more sort of bespoke tailored models. I think this is one that I’ll hand over to Hans to comment on.

Chih Han Chen   10:31

Thank you, Tom. So yes, so for for the if you’re really looking into this, like job with AI on public large data that says general purpose, then it’s spread out. And when you ask specific questions is looking responded in a certain way, that is from this experience, and it can be, it can be very general. And when you mentally train it, which is in Virtual Science, here, we have three levels of training, one is built using the medical life medical literature’s to try to that starting the, to have the base understanding of English or any other languages, from the perspective of medical, so then when it responds for the output, any information is related to medical descriptions, and then this, and then we take it to the next level. So we also tune it to a domain. So when we work with the clients, we understand that they’re working with, for example, HIV, that that disease area, we will gather that specific information to feed it to tune it further down to the domain, whatever domain, it can be, backdrop, it can be a treatment, it can be disease, finding or discussions. And then finally, we have the third layer, which is when the client works with us, we actually use their interactions data to full training that because they may be they want to invent a new word that never appeared anywhere else, they invent a new drug name, they invent a new pathway, that that actually requires further training. And then we isolate that that data, that model in a compliant isolated environment, that only service is popular clients. That’s, that’s how we differentiate ourselves to charging.

Garth Sundem   12:38

Right, although what I pointed out here is that, you know, if GPT has trained any on the internet, any earlier than six months ago, the field has changed so much for our companies that we need to be applying for the training to it. But you know, another another pain point that I hear from a lot of data scientists is that we always used to have to have one data set, we used to have to have data sort of healed together at the level of the data itself. Geraldine, is it true that now these Gen AI AIS can look across fragmented data and find meaning, not just in this one clean data set that we presented? But that is sets of differing purpose, different quality, different source, different formatting? Is that true?

Geraldine Reilly   13:40

Yes, absolutely. It’s true. And I’d say the thing that we are witnessing, companies that we’re collaborating with saying it, we really are hoping them to have a much more vision around cultivating collaboration within the organization, because the platform and the analytics are able to take that those datasets that you refer to those diverse data sets, preventing them from being viewed in silos within the organizations, and being able to be applied for so effectively creating a much better cross functional way of working. Ultimately, this speeds up access to the information that ultimately enables companies to bring their drugs to market faster to meet unmet medical needs, which is really the applicability of what we’re talking about for patients outcomes.

Garth Sundem   14:41

Oh, no, that’s interesting. I was thinking about fragmented data sets being things like, you know, generating new knowledge from a patient registry and also electronic health records. And you’re saying that the same approach to a Breaking down data silos can be used internally within the organization to take a data set from commercial and take a data set from CLIN Dev and take a data set from Medical Affairs, and make company wide conclusions based on internal diverse data sets.

Geraldine Reilly   15:19

Quite right. And as we know, the time to approval of your, and the license indication that you get depends on that diversity. So if you are only working with a small group of patients, you will get a very restricted label, if you don’t recruit into your studies most rapidly enough, it’s gonna impact on your entire timeline. So now you’re taking external and internal knowledge. And you’re applying it completely across the board, looking at having a global view of in an organization, as opposed to a siloed view of the functional wherever you’re working in, and say, what else do we need to be working with? How else can we get a bigger, diverse group of voice into our studies?

Garth Sundem   16:07

Okay, so the more data, the more data diversity, externally and internally leads to faster and more robust approvals?

Geraldine Reilly   16:20

That’s absolutely right. We’ve been working with a company just recently, who had an unfavorable hearing from the FDA, and we’ve been working with them very closely, to rapidly absorbed various data sets, so that they’ve been able to respond in record time back to the FDA, with regard to their response documents. And that’s, that’s real life, that’s what’s happening in the field.

Garth Sundem   16:46

So Tom, this sounds like it goes beyond adboards. Are is the same, or the same types of Gen AI technologies, able to find insights in platforms or data for which they weren’t specifically designed? I, I mean, your technology caught my eye was designed for adboards. But now we’re talking about diverse data sources? Are we having to reinvent the wheel? Or? Or can we use the same sorts of technologies?

Tom Hughes   17:21

You know, it’s, it’s, you know, we’re really pleased about the direction that we’re going in. So when we started the company, you know, we’ve set out to become the market leader for advisory panel solutions, right, we’ve got a, an overtime engagement platform that is AI enabled to produce Insight Analytics that have never been produced before following these activities. What we were hearing from our clients is that they were like, this is cool. And, you know, this is, you know, it’s fast, they’re getting new insights. But what they were saying to us was that your AI could be applied to other datasets. And could you, you know, make sense of these other datasets. And, you know, I’ll give you an example. So, you know, we’re working in oncology, and, you know, it’s a fast evolving space that we’re working in, within oncology. And, you know, there’s a particular team that, you know, wants the better understand that disease area as it as it evolves. And it’s a medical adverse team. And, you know, one of the challenges that they’ve faced is that there’s a lot of data sets, whether that’s social or internal, whoever’s MSL data that disconnected and they’re looking for insights in this data. And you know, really what they want us to do. And what we’re working with them to do is to start to bring together those datasets and contextualize them in a way that makes it very easy to understand and aggregates the findings so they can look at the trends over time. And it was just something that came up through collaborating with our clients. And so we started to go in this direction.

Garth Sundem   19:06

So Tom, Geraldine, we have these wonderful insights. They come from all these data sources, we use our Gen AI, it spits out these learnings that we hope are actionable, we take them to our leadership, and the leadership says really? Gen AI? You know, how are we supposed to believe this? Or or or variations on that theme? So what would you say to Medical Affairs teams that are trying to communicate the impact of the insights generated by Gen AI?

Tom Hughes   19:40

I mean, like, let me start by saying that we do have a human in the loop. So whilst we are applying artificial intelligence, it’s, you know, 80% ai and 20% human to support with the QA QC. So you know, you can have very high confidence that the data that’s been and produce this is getting a review and it’s referenceable. So and then in terms of communicating to leadership, perhaps you’re the you want to make a comment on that.

Geraldine Reilly   20:13

Yeah, I mean, the buzzword, I think until I started being the buzzword was omni channel. And what we see here is you addressing omni channel, this is how the customers want to interact with the pharmaceutical companies, they want to work across the board with them, they want to work in new ways, the customers are really excited about working with AI, as well. They’re not frightened of it whatsoever. They love this idea that a synchronous platform where they can work over a time that suits them. So you really starting to address what the customer really wants, as opposed to what the pharma company thinks that they want. And that really is making a huge difference and senior management are really listening to that they really love to hear the fact that you’re engaging your customers in a way that the customers want to be engaged with. And ultimately, that gives you better insights, more actionable insights more quickly, in more depth in a time that really is sustainable for everybody.

Garth Sundem   21:11

That’s interesting. So Gen AI offers customer centricity in some of the same ways that omni channel would, oh, I don’t want to take us down the rabbit hole of comparing omni channel to Gen AI and how they would all work together because I know that there’s some synergies and also some competitions in those areas. Let’s let’s leave it at that for today. So thanks, Tom, Hans, and Gerald for joining us. To learn more about how your organization can partner with Virtual Science AI visit Virtual Science ai.com MAPS members don’t forget to subscribe. And we hope you enjoyed this episode of the Medical Affairs Professional Society podcast series: “Elevate”.